{
"github_url":"https://github.com/sophgo/tdl_models/tree/main/",
"_comment":"model_list specify the model maintained in tdl_models,each model should at least have file name and rgb_order(choice:rgb,bgr,gray)",
"model_list":{
"MBV2_DET_PERSON":{
"types":["person"],
"file_name":"mbv2_det_person_256_448_INT8",
"rgb_order":"rgb"
},
"YOLOV8N_DET_HAND":{
"types":["hand"],
"file_name":"yolov8n_det_hand_384_640_INT8"
},
"YOLOV8N_DET_PET_PERSON":{
"types":["cat","dog","person"],
"file_name":"yolov8n_det_pet_person_384_640_INT8"
},
"YOLOV8N_DET_BICYCLE_MOTOR_EBICYCLE":{
"types":["bicycle","motorcycle","ebicycle"],
"file_name":"yolov8n_det_bicycle_motor_ebicycle_384_640_INT8"
},
"YOLOV8N_DET_PERSON_VEHICLE":{
"types":["car","bus","truck","rider with motorcycle","person","bike","motorcycle"],
"file_name":"yolov8n_det_person_vehicle_384_640_INT8"
},
"YOLOV8N_DET_HAND_FACE_PERSON":{
"types":["hand","face","person"],
"file_name":"yolov8n_det_hand_face_person_384_640_INT8"
},
"YOLOV8N_DET_FACE_HEAD_PERSON_PET":{
"types":["face","head","person","pet"],
"file_name":"yolov8n_det_face_head_person_pet_384_640_INT8"
},
"YOLOV8N_DET_HEAD_PERSON":{
"types":["head","person"],
"file_name":"yolov8n_det_head_person_384_640_INT8"
},
"YOLOV8N_DET_HEAD_HARDHAT":{
"types":["head","hardhat"],
"file_name":"yolov8n_det_head_hardhat_576_960_INT8"
},
"YOLOV8N_DET_FIRE_SMOKE":{
"types":["fire","smoke"],
"file_name":"yolov8n_det_fire_smoke_384_640_INT8"
},
"YOLOV8N_DET_FIRE":{
"types":["fire"],
"file_name":"yolov8n_det_fire_384_640_INT8"
},
"YOLOV8N_DET_HEAD_SHOULDER":{
"types":["head shoulder"],
"file_name":"yolov8n_det_head_shoulder_384_640_INT8"
},
"YOLOV8N_DET_LICENSE_PLATE":{
"types":["license plate"],
"file_name":"yolov8n_det_license_plate_384_640_INT8"
},
"YOLOV8N_DET_TRAFFIC_LIGHT":{
"types":["red","yellow","green","off","wait on"],
"file_name":"yolov8n_det_traffic_light_384_640_INT8"
},
"YOLOV8N_DET_MONITOR_PERSON":{
"types":["person"],
"file_name":"yolov8n_det_monitor_person_256_448_INT8"
},
"YOLOV11N_DET_MONITOR_PERSON":{
"types":["person"],
"file_name":"yolov11n_det_monitor_person_384_640_INT8"
},
"YOLOV11N_DET_BICYCLE_MOTOR_EBICYCLE":{
"types":["bicycle","motorcycle","ebicycle"],
"file_name":"yolov11n_det_bicycle_motor_ebicycle_384_640_INT8"
},
"YOLOV5_DET_COCO80":{
"is_coco_types":true,
"file_name":"yolov5s_det_coco80_640_640_INT8"
},
"YOLOV6_DET_COCO80":{
"is_coco_types":true,
"file_name":"yolov6n_det_coco80_640_640_INT8"
},
"YOLOV7_DET_COCO80":{
"is_coco_types":true,
"file_name":"yolov7_tiny_det_coco80_640_640_INT8"
},
"YOLOV8_DET_COCO80":{
"is_coco_types":true,
"file_name":"yolov8n_det_coco80_640_640_INT8"
},
"YOLOV10_DET_COCO80":{
"is_coco_types":true,
"file_name":"yolov10n_det_coco80_640_640_INT8"
},
"YOLOV11N_DET_COCO80":{
"is_coco_types":true,
"file_name":"yolov11n_det_coco80_640_640_INT8"
},
"PPYOLOE_DET_COCO80":{
"is_coco_types":true,
"file_name":"ppyoloe_det_coco80_640_640_INT8"
},
"YOLOX_DET_COCO80":{
"is_coco_types":true,
"file_name":"yolox_m_det_coco80_640_640_INT8"
},
"YOLOV5":{
"_comment":"custom model, specify num_cls or branch string",
"file_name":""
},
"YOLOV6":{
"_comment":"custom model, specify num_cls or branch string",
"file_name":""
},
"YOLOV8":{
"_comment":"custom model, specify num_cls or branch string",
"file_name":"yolov8n_cv181x_int8_sym"
},
"YOLOV10":{
"_comment":"custom model, specify num_cls or branch string",
"file_name":""
},
"PPYOLOE":{
"_comment":"custom model, specify num_cls or branch string",
"file_name":""
},
"YOLOX":{
"_comment":"custom model, specify num_cls or branch string",
"file_name":""
},
"SCRFD_DET_FACE":{
"_comment":"output face and 5 landmarks",
"types":["face"],
"file_name":"scrfd_det_face_432_768_INT8"
},
"CLS_ATTRIBUTE_GENDER_AGE_GLASS":{
"_comment":"output age,gender(0:male,1:female),glass(0:no glass,1:glass)",
"types":["age","gender","glass"],
"file_name":"cls_attribute_gender_age_glass_112_112_INT8",
"rgb_order":"rgb",
"mean":[0,0,0],
"std":[255.0,255.0,255.0]
},
"CLS_ATTRIBUTE_GENDER_AGE_GLASS_MASK":{
"_comment":"output age,gender(0:male,1:female),glass(0:no glass,1:glass),mask(0:no mask,1:mask)",
"types":["age","gender","glass","mask"],
"file_name":"cls_attribute_gender_age_glass_mask_112_112_INT8",
"rgb_order":"rgb",
"mean":[0,0,0],
"std":[255.0,255.0,255.0]
},
"CLS_ATTRIBUTE_GENDER_AGE_GLASS_EMOTION":{
"_comment":"output age,gender(0:male,1:female),glass(0:no glass,1:glass),emotion(0:anger,1:disgut,2:fear,3:happy,4:neutral,5:sad;6:surprise)",
"types":["age","gender","glass","emotion"],
"file_name":"cls_attribute_gender_age_glass_emotion_112_112_INT8",
"rgb_order":"rgb",
"mean":[0,0,0],
"std":[255.0,255.0,255.0]
},
"CLS_RGBLIVENESS":{
"_comment":"output 0:live or 1:spoof",
"types":["live","spoof"],
"file_name":"cls_rgbliveness_256_256_INT8",
"rgb_order":"rgb",
"mean":[0,0,0],
"std":[255.0,255.0,255.0]
},
"CLS_YOLOV8":{
"file_name":"yolov8n_cv181x_int8_sym",
"rgb_order":"rgb",
"mean":[0,0,0],
"std":[255.0,255.0,255.0]
},
"CLS_HAND_GESTURE":{
"_comment":"output hand gesture(0:fist,1:five,2:none,3:two)",
"types":["fist","five","none","two"],
"file_name":"cls_hand_gesture_128_128_INT8",
"rgb_order":"rgb",
"mean":[0,0,0],
"std":[255.0,255.0,255.0]
},
"CLS_KEYPOINT_HAND_GESTURE":{
"_comment":"output hand gesture(0:fist,1:five,2:four,3:none,4:ok,5:one,6:three,7:three2,8:two)",
"types":["fist","five","four","none","ok","one","three","three2","two"],
"file_name":"cls_keypoint_hand_gesture_1_42_INT8",
"rgb_order":"rgb",
"mean":[0,0,0],
"std":[1.0,1.0,1.0]
},
"CLS_SOUND_BABAY_CRY":{
"_comment":"output 0:background or 1:cry,single channel",
"types":["background","cry"],
"file_name":"cls_sound_babay_cry_188_40_INT8",
"rgb_order":"gray"
},
"CLS_SOUND_COMMAND_NIHAOSHIYUN":{
"_comment":"single channel,TODO:add types",
"types":["background","nihaoshiyun"],
"file_name":"cls_sound_nihaoshiyun_126_40_INT8",
"rgb_order":"gray",
"hop_len":128,
"fix":1
},
"CLS_SOUND_COMMAND_NIHAOSUANNENG":{
"_comment":"single channel,TODO:add types",
"types":["background","nihaosuanneng"],
"file_name":"my_custom_sound_command",
"rgb_order":"gray",
"hop_len":128,
"fix":1
},
"CLS_SOUND_COMMAND_XIAOAIXIAOAI":{
"_comment":"single channel,TODO:add types",
"types":["background","xiaoaixiaoai"],
"file_name":"cls_sound_xiaoaixiaoai_126_40_INT8",
"rgb_order":"gray",
"hop_len":128,
"fix":1
},
"CLS_IMG":{
"_comment":"custom classification, set types,file_name,specify rgb order and mean/std",
"types":["custom"],
"file_name":"",
"rgb_order":"rgb"
},
"KEYPOINT_LICENSE_PLATE":{
"_comment":"output 4 license plate keypoints",
"types":["top_left","top_right","bottom_left","bottom_right"],
"file_name":"keypoint_license_plate_64_128_INT8",
"rgb_order":"rgb"
},
"KEYPOINT_HAND":{
"_comment":"output 21 hand keypoints",
"file_name":"keypoint_hand_128_128_INT8",
"rgb_order":"rgb"
},
"KEYPOINT_YOLOV8POSE_PERSON17":{
"_comment":"output 17 person keypoints and box",
"file_name":"keypoint_yolov8pose_person17_384_640_INT8",
"rgb_order":"rgb"
},
"KEYPOINT_SIMCC_PERSON17":{
"_comment":"output 17 person keypoints from cropped image",
"file_name":"keypoint_simcc_person17_256_192_INT8",
"rgb_order":"rgb"
},
"KEYPOINT_FACE_V2": {
"_comment": "KEYPOINT_FACE_V2",
"file_name": "keypoint_face_v2_64_64_INT8"
},
"LSTR_DET_LANE":{
"_comment":"output lane keypoints",
"file_name":"lstr_det_lane_360_640_MIX",
"rgb_order":"rgb"
},
"RECOGNITION_LICENSE_PLATE":{
"_comment":"output 7 license plate characters",
"file_name":"recognition_license_plate_24_96_MIX",
"rgb_order":"bgr"
},
"YOLOV8_SEG":{
"_comment":"custom segmentation,set types,file_name,specify rgb order",
"types":[],
"file_name":"yolov8_seg_384_640_INT8"
},
"YOLOV8_SEG_COCO80":{
"is_coco_types":true,
"_comment":"output 80 segmentation mask",
"file_name":"yolov8n_seg_coco80_640_640_INT8"
},
"TOPFORMER_SEG_PERSON_FACE_VEHICLE":{
"_comment":"output mask",
"types":["background","person","face","vehicle","license plate"],
"file_name":"topformer_seg_person_face_vehicle_384_640_INT8",
"rgb_order":"rgb"
},
"FEATURE_IMG":{
"_comment":"custom segmentation,set file_name,specify rgb order,set mean/std",
"file_name":"",
"rgb_order":"rgb"
},
"FEATURE_CLIP_IMG":{
"_comment":"clip image feature extraction",
"file_name":"feature_clip_image_224_224_W4BF16",
"rgb_order":"rgb"
},
"FEATURE_CLIP_TEXT":{
"_comment":"clip text feature extraction",
"file_name":"feature_clip_text_1_77_W4BF16",
"rgb_order":"rgb"
},
"FEATURE_MOBILECLIP2_IMG":{
"_comment":"mobileclip2 image feature extraction",
"file_name":"feature_mobileclip2_B_img_224_224_INT8",
"rgb_order":"rgb"
},
"FEATURE_MOBILECLIP2_TEXT":{
"_comment":"mobileclip2 text feature extraction",
"file_name":"feature_mobileclip2_B_text_1_77_INT8",
"rgb_order":"rgb"
},
"FEATURE_CVIFACE":{
"_comment":"cviface 256-dimensional feature",
"file_name":"feature_cviface_112_112_INT8",
"rgb_order":"rgb",
"mean":[127.5,127.5,127.5],
"std":[128,128,128]
},
"FEATURE_BMFACE_R34":{
"_comment":"output 512 dim feature",
"file_name":"feature_bmface_r34_112_112_INT8",
"rgb_order":"rgb",
"mean":[0,0,0],
"std":[1,1,1]
},
"FEATURE_BMFACE_R50":{
"_comment":"output 512 dim feature",
"file_name":"bmface_r50_v1_bmnetp.bmodel",
"rgb_order":"rgb",
"mean":[0,0,0],
"std":[1,1,1]
},
"TRACKING_FEARTRACK":{
"_comment":"single object tracking",
"file_name":"tracking_feartrack_128_128_256_256_INT8",
"rgb_order":"rgb",
"mean":[123.675,116.28,103.53],
"std":[58.395,57.12,57.375]
},
"RECOGNITION_SPEECH_ZIPFORMER_ENCODER":{
"file_name":"recognition_speech_zipformer_encoder-s_71_80_BF16"
},
"RECOGNITION_SPEECH_ZIPFORMER_DECODER":{
"file_name":"recognition_speech_zipformer_decoder-s_1_2_BF16"
},
"RECOGNITION_SPEECH_ZIPFORMER_JOINER":{
"file_name":"recognition_speech_zipformer_joiner-s_1_512_1_512_BF16"
}
}
}
=================
!!!!!!!!!!!!!!!!!!!!!!!! solucao tdl_sdk !!!!!!!!!!!!!!!!!!!!!!!
em
https://colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/train-yolov8-obb.ipynb#scrollTo=tdSMcABDNKW-
!pip install ultralytics==8.2.103 -q
!pip install ultralytics==8.3.0 #o segredo para o yolo nao pegar .pt superior
treine
from ultralytics import YOLO
model = YOLO('yolov8n.pt')
results = model.train(data=f"data.yaml", epochs=100, imgsz=640)
https://github.com/milkv-duo/duo-buildroot-sdk-v2/blob/develop/tdl_sdk/tool/yolo_export/yolov8_export.py
model_transform \
--model_name yolov8n \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0.0,0.0,0.0 \
--scale 0.0039216,0.0039216,0.0039216 \
--keep_aspect_ratio \
--pixel_format rgb \
--mlir yolov8n.mlir
fotos na pasta BUGGIO estavam 640x640
run_calibration.py yolov8n.mlir \
--dataset ../../BUGGIO \
--input_num 100 \
-o yolov8n_cali_table
model_deploy.py \
--mlir yolov8n.mlir \
--quant_input --quant_output \
--quantize INT8 \
--calibration_table yolov8n_cali_table \
--processor cv181x \
--model yolov8n_cv181x_int8_sym.cvimodel
python tdl sdk sophgo example
tdl_sophgo.py
python tdl_sophgo.py /root/cv181x/yolov8n_cv181x_int8_sym_cv181x.cvimodel Barbie_5-4_jpg.rf.64feb144416c82dc7c58b335d7143774.jpg
import sys
import os
from tdl import nn, image
import cv2
import numpy as np
if __name__ == "__main__":
if len(sys.argv) != 3:
print("Usage: python sample_fd.py <model_path> <image_path>")
sys.exit(1)
model_path = sys.argv[1]
img_path = sys.argv[2]
face_detector = nn.get_model(nn.ModelType.YOLOV8, model_path)
img = image.read(img_path)
# img = cv2.imread(img_path)
bboxes = face_detector.inference(img)
print(bboxes)
#https://github.com/sophgo/tdl_sdk/tree/master
python tdl sdk milkv-duo S example
sample_img_object_detection.py
python sample_img_object_detection.py YOLOV8 /root/cv181x/yolov8n_cv181x_int8_sym_cv181x.cvimodel /root/Bauer_9-4_jpg.rf.1ee8c79f82e5c4ed6b2ba3b7d5340d2c.jpg
import sys
import os
from tdl import nn, image
import cv2
import numpy as np
def visualize_objects(img_path, bboxes, save_path="object_detection.jpg"):
"""å¯è§†åŒ–ç›®æ ‡æ£€æµ‹ç»“æžœ"""
img = cv2.imread(img_path)
print(f"检测到 {len(bboxes)} ä¸ªç›®æ ‡")
for i, obj in enumerate(bboxes):
x1, y1, x2, y2 = map(int, [obj['x1'], obj['y1'], obj['x2'], obj['y2']])
class_id = obj['class_id']
score = obj['score']
class_name = obj.get('class_name', f'class_{class_id}')
cv2.rectangle(img, (x1, y1), (x2, y2), (0, 0, 255), 2)
label = f"{class_id}:{score:.2f}"
center_x = (x1 + x2) // 2
center_y = (y1 + y2) // 2
cv2.putText(img, label, (center_x, center_y),
cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1)
cv2.imwrite(save_path, img)
print(f"ä¿å˜å›¾åƒåˆ°: {save_path}")
if __name__ == "__main__":
if len(sys.argv) < 4 or len(sys.argv) > 5:
print("Usage: python3 sample_img_object_detection.py <model_id_name> <model_dir> <image_path> [threshold]")
# å°šæœªåŠ å…¥å¯¹äºŽä¸æ˜¯æ£€æµ‹æ¨¡åž‹çš„处ç†
sys.exit(1)
model_id_name = sys.argv[1]
model_dir = sys.argv[2]
image_path = sys.argv[3]
threshold = float(sys.argv[4]) if len(sys.argv) == 5 else 0.5
if not os.path.exists(image_path):
print(f"图åƒæ–‡ä»¶ä¸å˜åœ¨: {image_path}")
sys.exit(1)
model_type = getattr(nn.ModelType, model_id_name)
model = nn.get_model(model_type, model_dir, device_id=0)
# 读å–图åƒ
img = image.read(image_path)
# 执行推ç†
outdatas = model.inference(img)
expected_keys = {"class_id", "class_name", "score", "x1", "y1", "x2", "y2"}
is_detection = (
isinstance(outdatas, list) and
isinstance(outdatas[0], dict) and
set(outdatas[0].keys()) == expected_keys
)
if not is_detection:
print("当å‰æ¨¡åž‹ä¸æ˜¯ç›®æ ‡æ£€æµ‹æ¨¡åž‹ï¼Œè¾“出内容：")
print(outdatas)
sys.exit(1)
print(f"out_datas.size: {len(outdatas)}")
for i, obj in enumerate(outdatas):
print(f"obj_meta_index: {i} "
f"class: {obj['class_id']} "
f"score: {obj['score']:.2f} "
f"bbox: {obj['x1']:.2f} {obj['y1']:.2f} {obj['x2']:.2f} {obj['y2']:.2f}")
visualize_objects(image_path, outdatas)
# input: python3 sample_img_object_detection.py <model_id_name> <model_dir> <image_path> [threshold]
# output: obj_meta_index: <index> class: <class_id> score: <score_value> bbox: <x1> <y1> <x2> <y2>
utilizando Script completo
./pt_to_cvimodel_tdl_sdk.sh
python yolov8_export.py
python export_tdl_sdk.py --dataset ../../BUGGIO --test_input ../../BUGGIO/Ades_2-3_jpg.rf.de3d17a6dcc748c6642882198a1c1c76.jpg best.onnx
!!!!!!!!!!!!!!!!!!!!!!!! solucao recamera !!!!!!!!!!!!!!!!!!!!!!!
em
https://colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/train-yolo11-object-detection-on-custom-dataset.ipynb#scrollTo=tdSMcABDNKW-yolo11s
!pip install ultralytics==8.2.103 -q
faça o upgrade
pip install --upgrade --force-reinstall ultralytics
copie o zip do BUGGIO
!unzip /context/buggio.v1i.yolov11.zip
treine
<---> Python :)
from ultralytics import YOLO
model = YOLO("/content/yolo11n.pt")
model.train(data="data.yaml", epochs=200, imgsz=640)
./pt_to_cvimodel_recamera.sh
yolo export model=best.pt format=onnx imgsz=640,640
python export_recamera.py --output_names "/model.23/cv2.0/cv2.0.2/Conv_output_0,/model.23/cv3.0/cv3.0.2/Conv_output_0,/model.23/cv2.1/cv2.1.2/Conv_output_0,/model.23/cv3.1/cv3.1.2/Conv_output_0,/model.23/cv2.2/cv2.2.2/Conv_output_0,/model.23/cv3.2/cv3.2.2/Conv_output_0" --dataset ../../BUGGIO --test_input ../../BUGGIO/Barbie_7-10_jpg.rf.502fbff248ff3b2336a9e60317de843b.jpg best.onnx --quantize INT8
testar cvimodel no recamera
model_transform --model_name best --model_def ./best.onnx --input_shapes '[[1,3,640,640]]' --mean 0.0,0.0,0.0 --scale 0.0039216,0.0039216,0.0039216 --keep_aspect_ratio --pixel_format rgb --output_names /model.23/cv2.0/cv2.0.2/Conv_output_0,/model.23/cv3.0/cv3.0.2/Conv_output_0,/model.23/cv2.1/cv2.1.2/Conv_output_0,/model.23/cv3.1/cv3.1.2/Conv_output_0,/model.23/cv2.2/cv2.2.2/Conv_output_0,/model.23/cv3.2/cv3.2.2/Conv_output_0 --test_input ../../BUGGIO/Ades_2-4_jpg.rf.4de8403c125c5d16b435a839a3a93780.jpg --test_result best_top_outputs.npz --mlir best.mlir
run_calibration best.mlir --dataset ../../BUGGIO --input_num 200 -o best_calib_table
model_deploy --mlir best.mlir --quantize INT8 --quant_input --processor cv181x --calibration_table best_calib_table --test_input ../../BUGGIO/Ades_2-4_jpg.rf.4de8403c125c5d16b435a839a3a93780.jpg --test_reference best_top_outputs.npz --customization_format RGB_PACKED --fuse_preprocess --aligned_input --model best_cv181x_int8.cvimodel
model_transform \
--model_name yolo11n \
--model_def best.onnx \
--input_shapes "[[1,3,640,640]]" \
--mean "0.0,0.0,0.0" \
--scale "0.0039216,0.0039216,0.0039216" \
--keep_aspect_ratio \
--pixel_format rgb \
--output_names "/model.23/cv2.0/cv2.0.2/Conv_output_0,/model.23/cv3.0/cv3.0.2/Conv_output_0,/model.23/cv2.1/cv2.1.2/Conv_output_0,/model.23/cv3.1/cv3.1.2/Conv_output_0,/model.23/cv2.2/cv2.2.2/Conv_output_0,/model.23/cv3.2/cv3.2.2/Conv_output_0" \
--test_input ../../BUGGIO/Ades_2-3_jpg.rf.de3d17a6dcc748c6642882198a1c1c76.jpg \
--test_result yolo11n_top_outputs.npz \
--mlir yolo11n.mlir
run_calibration \
yolo11n.mlir \
--dataset ../../BUGGIO \
--input_num 100 \
-o yolo11n_calib_table
model_deploy \
--mlir yolo11n.mlir \
--quantize INT8 \
--quant_input \
--processor cv181x \
--calibration_table yolo11n_calib_table \
--test_input ../../BUGGIO/Ades_2-3_jpg.rf.de3d17a6dcc748c6642882198a1c1c76.jpg \
--test_reference yolo11n_top_outputs.npz \
--customization_format RGB_PACKED \
--fuse_preprocess \
--aligned_input \
--model yolo11n_1684x_int8_sym.cvimodel
=======car counting 11
https://github.com/Seeed-Studio/sscma-example-sg200x/tree/main/solutions/sscma-model/main
yolo export model=best.pt format=onnx opset=14
python export_recamera.py --output_names "/model.23/cv2.0/cv2.0.2/Conv_output_0,/model.23/cv3.0/cv3.0.2/Conv_output_0,/model.23/cv2.1/cv2.1.2/Conv_output_0,/model.23/cv3.1/cv3.1.2/Conv_output_0,/model.23/cv2.2/cv2.2.2/Conv_output_0,/model.23/cv3.2/cv3.2.2/Conv_output_0" --dataset ../../CARS --test_input ../../CARS/DOH_3-video-converter_com-_mp4-26_jpg.rf.a6a631199f4152b1ab619e3e3cf6e8ee.jpg best.onnx --quantize INT8
model_transform \
--model_name yolo11n \
--model_def best.onnx \
--input_shapes "[[1,3,640,640]]" \
--mean "0.0,0.0,0.0" \
--scale "0.0039216,0.0039216,0.0039216" \
--keep_aspect_ratio \
--pixel_format rgb \
--output_names "/model.23/cv2.0/cv2.0.2/Conv_output_0,/model.23/cv3.0/cv3.0.2/Conv_output_0,/model.23/cv2.1/cv2.1.2/Conv_output_0,/model.23/cv3.1/cv3.1.2/Conv_output_0,/model.23/cv2.2/cv2.2.2/Conv_output_0,/model.23/cv3.2/cv3.2.2/Conv_output_0" \
--test_input ../../CARS/DOH_3-video-converter_com-_mp4-26_jpg.rf.a6a631199f4152b1ab619e3e3cf6e8ee.jpg \
--test_result yolo11n_top_outputs.npz \
--mlir yolo11n.mlir
run_calibration \
yolo11n.mlir \
--dataset ../../CARS \
--input_num 100 \
-o yolo11n_calib_table
model_deploy \
--mlir yolo11n.mlir \
--quantize INT8 \
--quant_input \
--processor cv181x \
--calibration_table yolo11n_calib_table \
--test_input ../../CARS/DOH_3-video-converter_com-_mp4-26_jpg.rf.a6a631199f4152b1ab619e3e3cf6e8ee.jpg \
--test_reference yolo11n_top_outputs.npz \
--customization_format RGB_PACKED \
--fuse_preprocess \
--aligned_input \
--model yolo11n_1684x_int8_sym.cvimodel
yolo26n
model_transform \
--model_name yolo26n \
--model_def best_26.onnx \
--input_shapes "[[1,3,640,640]]" \
--mean "0.0,0.0,0.0" \
--scale "0.0039216,0.0039216,0.0039216" \
--keep_aspect_ratio \
--pixel_format rgb \
--output_names "/model.23/cv2.0/cv2.0.2/Conv_output_0,/model.23/cv3.0/cv3.0.2/Conv_output_0,/model.23/cv2.1/cv2.1.2/Conv_output_0,/model.23/cv3.1/cv3.1.2/Conv_output_0,/model.23/cv2.2/cv2.2.2/Conv_output_0,/model.23/cv3.2/cv3.2.2/Conv_output_0" \
--test_input ../../BUGGIO/Ades_2-4_jpg.rf.4de8403c125c5d16b435a839a3a93780.jpg \
--test_result yolo26n_top_outputs.npz \
--mlir yolo26n.mlir
run_calibration \
yolo26n.mlir \
--dataset ../../BUGGIO \
--input_num 100 \
-o yolo26n_calib_table
model_deploy \
--mlir yolo26n.mlir \
--quantize INT8 \
--quant_input \
--processor cv181x \
--calibration_table yolo26n_calib_table \
--test_input ../../BUGGIO/Ades_2-4_jpg.rf.4de8403c125c5d16b435a839a3a93780.jpg \
--test_reference yolo26n_top_outputs.npz \
--customization_format RGB_PACKED \
--fuse_preprocess \
--aligned_input \
--model yolo26n_1684x_int8_sym.cvimodel
https://github.com/ultralytics/ultralytics/blob/ee2ac9e43491e5ca61a158fb3a42e621a6710ee1/docs/en/integrations/seeedstudio-recamera.md?plain=1#L62
https://github.com/Seeed-Studio/SSCMA-Micro
https://docs.ultralytics.com/modes/predict/#key-features-of-predict-mode
https://github.com/Seeed-Studio/reCamera-OS/tree/sg200x-reCamera/external/br2-external
ULTRALYTICS
OPEN COLAB
https://colab.research.google.com/github/ultralytics/ultralytics/blob/main/examples/tutorial.ipynb
Install Ultralytics
!uv pip install ultralytics
import ultralytics
ultralytics.checks()
copy the BUGGIO DATASET (got with ROBOFLOW)
!unzip /cont
yolo check
Train using yolo26n.pt reference using the data.yaml of roboblox
!yolo train model=yolo26n.pt data=data.yaml epochs=100 imgsz=640
!yolo classify train data=/content/train model=yolo26n-cls.pt epochs=100 imgsz=640
train.zip (class)
yolo export model=best.pt format=onnx opset=14 end2end=False
model_transform \
--model_name yolo26n \
--model_def best.onnx \
--input_shapes "[[1,3,640,640]]" \
--mean "0.0,0.0,0.0" \
--scale "0.0039216,0.0039216,0.0039216" \
--keep_aspect_ratio \
--pixel_format rgb \
--output_names "/model.23/cv2.0/cv2.0.2/Conv_output_0,/model.23/cv3.0/cv3.0.2/Conv_output_0,/model.23/cv2.1/cv2.1.2/Conv_output_0,/model.23/cv3.1/cv3.1.2/Conv_output_0,/model.23/cv2.2/cv2.2.2/Conv_output_0,/model.23/cv3.2/cv3.2.2/Conv_output_0" \
--test_input ../../BUGGIO/Ades_2-4_jpg.rf.4de8403c125c5d16b435a839a3a93780.jpg \
--test_result yolo26n_top_outputs.npz \
--mlir yolo26n.mlir > transform.txt
run_calibration \
yolo26n.mlir \
--dataset ../../BUGGIO \
--input_num 100 \
-o yolo26n_calib_table > calib.txt
model_deploy \
--mlir yolo26n.mlir \
--quantize INT8 \
--quant_input \
--processor cv181x \
--calibration_table yolo26n_calib_table \
--test_input ../../BUGGIO/Ades_2-4_jpg.rf.4de8403c125c5d16b435a839a3a93780.jpg \
--test_reference yolo26n_top_outputs.npz \
--customization_format RGB_PACKED \
--fuse_preprocess \
--aligned_input \
--model yolo26n_1684x_int8_sym.cvimodel > cvimodel.txt
yolo export model=best.pt format=onnx opset=14 end2end=False
yolo export model=best.pt format=onnx opset=14
pip install --upgrade pip
pip install git+https://github.com/ultralytics/ultralytics.git@main
yolo export model=best.pt format=onnx opset=14 end2end=False
model_transform \
--model_name yolo26n \
--model_def yolo26n.onnx \
--input_shapes "[[1,3,640,640]]" \
--mean "0.0,0.0,0.0" \
--scale "0.0039216,0.0039216,0.0039216" \
--keep_aspect_ratio \
--pixel_format rgb \
--test_input ../../BUGGIO/Ades_2-4_jpg.rf.4de8403c125c5d16b435a839a3a93780.jpg \
--test_result yolo26n_top_outputs.npz \
--mlir yolo26n.mlir
run_calibration \
yolo26n.mlir \
--dataset ../../BUGGIO \
--input_num 100 \
-o yolo26n_calib_table
model_deploy \
--mlir yolo26n.mlir \
--quantize INT8 \
--quant_input \
--processor cv181x \
--calibration_table yolo26n_calib_table \
--test_input ../../BUGGIO/Ades_2-4_jpg.rf.4de8403c125c5d16b435a839a3a93780.jpg \
--test_reference yolo26n_top_outputs.npz \
--customization_format RGB_PACKED \
--fuse_preprocess \
--aligned_input \
--model yolo26n_1684x_int8_sym.cvimodel
yolo export model=yolov26n.pt format=onnx opset=14 end2end=False
2.8.12
cmake
===========desespero
yolo export model=best.pt format=onnx opset=14 end2end=False
model_transform \
--model_name yolo11n \
--model_def best.onnx \
--input_shapes "[[1,3,640,640]]" \
--mean "0.0,0.0,0.0" \
--scale "0.0039216,0.0039216,0.0039216" \
--keep_aspect_ratio \
--pixel_format rgb \
--output_names "/model.23/cv2.0/cv2.0.2/Conv_output_0,/model.23/cv3.0/cv3.0.2/Conv_output_0,/model.23/cv2.1/cv2.1.2/Conv_output_0,/model.23/cv3.1/cv3.1.2/Conv_output_0,/model.23/cv2.2/cv2.2.2/Conv_output_0,/model.23/cv3.2/cv3.2.2/Conv_output_0" \
--test_input ../image/dog.jpg \
--test_result yolo11n_top_outputs.npz \
--mlir yolo11n.mlir
run_calibration \
yolo11n.mlir \
--dataset ../COCO2017 \
--input_num 100 \
-o yolo11n_calib_table
/workspace/duo-buildroot-sdk-v2/buildroot/utils/brmake -j4 -C /workspace/duo-buildroot-sdk-v2/buildroot
2026-01-29T05:40:25 >>> host-tar 1.35 Configuring
Done in 1min 08s (error code: 2)
make: *** [Makefile:664: br-rootfs-pack] Error 2
Error: Build board milkv-duos-musl-riscv64-emmc failed!
root@f09cd7f3870c:/workspace/duo-buildroot-sdk-v2#
https://github.com/smuraro/duo-buildroot-sdk-v2
sudo apt install -y pkg-config build-essential ninja-build automake autoconf libtool wget curl git gcc libssl-dev bc slib squashfs-tools android-sdk-libsparse-utils jq python3-distutils scons parallel tree python3-dev python3-pip device-tree-compiler ssh cpio fakeroot libncurses5 flex bison libncurses5-dev genext2fs rsync unzip dosfstools mtools tcl openssh-client cmake expect
sudo apt para o python
export FORCE_UNSAFE_CONFIGURE=1
export PATH=$(echo $PATH | tr -d '\n\t ')
sudo apt-get install python-is-python3
sudo apt install nodejs npm
sudo chown -R root:root /workspace/duo-buildroot-sdk-v2
./build.sh lunch
cd buildroot-2021.05/output/milkv-duo_musl_riscv64
make show-targets
make menuconfig
https://wiki.seeedstudio.com/recamera_model_conversion/
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
pip install git+https://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@github.com/tcpipchip/ultralytics.git@main
pip install git+https://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@github.com/ultralytics/ultralytics.git@main
claude Windows
arquiteturadecomputadores1@gmail.com
claude windows (ps1)
irm https://claude.ai/install.ps1 | iex
C:\Users\Miguel\.local\bin\claude.exe skills install git+https://github.com/Seeed-Studio/ai-skills
I'll guide you through converting your custom YOLO v26 ONNX model to CVIMODEL format.
Prerequisites
1. TPU-MLIR toolchain (Sophgo's conversion tool)
2. Your custom .onnx model file 3. Calibration dataset (~100-1000 images for INT8 quantization)
4. A test image for validation
Conversion Steps
1. Install TPU-MLIR
# Clone TPU-MLIR v1.7 (or latest)
git clone -b v1.7 --depth 1 https://github.com/sophgo/tpu-mlir.git
cd tpu-mlir
source ./envsetup.sh
./build.sh
2. Convert ONNX to MLIR (FP32)
model_transform \
--model_name your_model_name \
--model_def your_model.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0.0,0.0,0.0 \
--scale 0.0039216,0.0039216,0.0039216 \
--keep_aspect_ratio \
--pixel_format rgb \
--test_input test_image.jpg \
--test_result output_ref.npz \
--mlir your_model.mlir
Important: You may need to specify --output_names with your model's specific output layer names. You can find these by inspecting your ONNX
model.
3. Run Calibration (for INT8 quantization)
run_calibration your_model.mlir \
--dataset /path/to/calibration/images \
--input_num 100 \
-o calibration_table
4. Convert to CVIMODEL
# For INT8 (recommended for edge devices)
model_deploy \
--mlir your_model.mlir \
--quantize INT8 \
--calibration_table calibration_table \
--chip cv181x \
--test_input test_image.jpg \
--test_reference output_ref.npz \
--model your_model_int8.cvimodel
# For BF16 (higher precision)
model_deploy \
--mlir your_model.mlir \
--quantize BF16 \
--chip cv181x \
--test_input test_image.jpg \
--test_reference output_ref.npz \
--model your_model_bf16.cvimodel
Key Questions for Your Setup
1. What chip are you targeting? (cv180x, cv181x, cv182x, cv183x)
2. What's your input size? (e.g., 640x640, 416x416)
3. Do you know your model's output layer names? (needed for --output_names)
Finding Output Layer Names
If you need to inspect your ONNX model:
pip install netron
netron your_model.onnx
Or use Python:
import onnx
model = onnx.load("your_model.onnx")
for output in model.graph.output:
print(output.name)
Would you like me to help you with any specific step, or do you need help finding your model's output layer names?
Sources:
- https://wiki.seeedstudio.com/model_conversion_guide/
- https://github.com/sophgo/tpu-mlir
- https://wiki.seeedstudio.com/recamera_model_conversion/
- https://doc.sophgo.com/sdk-docs/v23.05.01/docs_latest_release/docs/tpu-mlir/quick_start_en/html/10_cv18xx_guide.html
> cv181x 640x640
● Perfect! Here are the specific commands for your cv181x chip with 640x640 input:
Step-by-Step Conversion for Your Setup
1. Setup TPU-MLIR Environment
git clone -b v1.7 --depth 1 https://github.com/sophgo/tpu-mlir.git
cd tpu-mlir
source ./envsetup.sh
cd ..
2. Inspect Your ONNX Model (to find output layer names)
First, let's check what output layers your YOLO v26 model has:
# Install onnx if needed
pip install onnx
# Create inspection script
python -c "
import onnx
model = onnx.load('your_model.onnx')
print('Input names:')
for inp in model.graph.input:
print(f' {inp.name}')
print('\nOutput names:')
for out in model.graph.output:
print(f' {out.name}')
"
3. Convert ONNX to MLIR
model_transform \
--model_name yolov26_custom \
--model_def your_model.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0.0,0.0,0.0 \
--scale 0.0039216,0.0039216,0.0039216 \
--keep_aspect_ratio \
--pixel_format rgb \
--test_input test_image.jpg \
--test_result yolov26_output_ref.npz \
--mlir yolov26_custom.mlir
Note: Add --output_names output1,output2,output3 if you have specific output layer names from step 2.
4. Run Calibration
run_calibration yolov26_custom.mlir \
--dataset ./calibration_images \
--input_num 100 \
-o yolov26_calibration_table
5. Generate CVIMODEL (INT8)
model_deploy \
--mlir yolov26_custom.mlir \
--quantize INT8 \
--calibration_table yolov26_calibration_table \
--chip cv181x \
--test_input test_image.jpg \
--test_reference yolov26_output_ref.npz \
--model yolov26_custom_cv181x_int8.cvimodel
What You Need Ready
1. ✓ Your .onnx file
2. ✓ Target chip: cv181x
3. ✓ Input size: 640x640
4. ⚠️ Calibration dataset (100+ images similar to your training data)
5. ⚠️ Test image (for validation)
Questions
1. Do you have your ONNX file in the current directory? If so, I can help you inspect it to find the output layer names.
2. Do you have calibration images ready? They should be representative of your inference data.
3. Are you running this on Linux/WSL or Windows? (TPU-MLIR requires Linux environment)
Would you like me to help you inspect your ONNX model or prepare the calibration dataset?
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────> inspect my model.onnx
───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────
? for shortcuts
pip install -U ultralytics
yolo export model=best.pt format=onnx imgsz=640 opset=14 simplify=True dynamic=False
power shell
wsl2 ubuntu 22.04 install
sudo apt-get update
sudo apt-get upgrade
sudo apt install -y pkg-config build-essential ninja-build automake autoconf libtool wget curl git gcc libssl-dev bc slib squashfs-tools android-sdk-libsparse-utils jq python3-distutils scons parallel tree python3-dev python3-pip device-tree-compiler ssh cpio fakeroot libncurses5 flex bison libncurses5-dev genext2fs rsync unzip dosfstools mtools tcl openssh-client cmake expect
execute docker windows desktop
instalei ultralytics no ubuntu 22.04
copiei scripts dentro do ubuntu 22.04
git clone https://github.com/Seeed-Studio/ai-skills/tree/main
dentro da pasta /home/ubuntu/ai-skills/skills/onnx-to-cvimodel/scripts copiei o .pt
e dentro de /home/ubuntu/ai-skills/skills/onnx-to-cvimodel/scripts criei uma pasta dataset e copiei imagens
pip install -U ultralytics para gerar o onnx
conversao de .pt para .onnx
yolo export model=best.pt format=onnx imgsz=640 opset=14 simplify=True dynamic=False
nano ~/.docker/config.json (no ubuntu 22.04)
{
}
pip install tpu-mlir
git clone https://github.com/sophgo/tpu-mlir.git
./convert_to_cvimodel.sh best.onnx dataset
=====================
object detect
can detect many objects at the same time
model_transform \
--model_name yolov26_custom \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0.0,0.0,0.0 \
--scale 0.0039216,0.0039216,0.0039216 \
--keep_aspect_ratio \
--pixel_format rgb \
--test_input teste.jpg \
--test_result yolov26_output_ref.npz \
--output_names "/model.23/one2one_cv2.0/one2one_cv2.0.2/Conv_output_0,/model.23/one2one_cv3.0/one2one_cv3.0.2/Conv_output_0,/model.23/one2one_cv2.1/one2one_cv2.1.2/Conv_output_0,/model.23/one2one_cv3.1/one2one_cv3.1.2/Conv_output_0,/model.23/one2one_cv2.2/one2one_cv2.2.2/Conv_output_0,/model.23/one2one_cv3.2/one2one_cv3.2.2/Conv_output_0" \
--mlir yolov26_custom.mlir
run_calibration yolov26_custom.mlir \
--dataset ./calibration_images \
--input_num 200 \
-o yolov26_calibration_table
model_deploy \
--mlir yolov26_custom.mlir \
--quantize INT8 \
--calibration_table yolov26_calibration_table \
--chip cv181x \
--test_input teste.jpg \
--test_reference yolov26_output_ref.npz \
--model yolov26_custom_cv181x_int8.cvimodel
name: output0
tensor: float32[1,300,6]
=================
best class pt
model_transform \
--model_name yolov26_custom \
--model_def best_class.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0.0,0.0,0.0 \
--scale 0.0039216,0.0039216,0.0039216 \
--keep_aspect_ratio \
--pixel_format rgb \
--test_input teste.jpg \
--test_result yolov26_output_ref.npz \
--output_names "output0" \
--mlir yolov26_custom.mlir
run_calibration yolov26_custom.mlir \
--dataset ./calibration_images \
--input_num 200 \
-o yolov26_calibration_table
model_deploy \
--mlir yolov26_custom.mlir \
--quantize INT8 \
--calibration_table yolov26_calibration_table \
--chip cv181x \
--test_input teste.jpg \
--test_reference yolov26_output_ref.npz \
--model yolov26_custom_cv181x_int8.cvimodel
========================
testing scripts virtual box ubuntu 22.0.4
sem update ou upgrade
instalei docker apenas
nao precisa docker desktop windows...
sudo chown -R ubuntu:ubuntu ai-skills/
sudo ./convert_to_cvimodel.sh best.onnx images
criei pasta compartilhada no virtual box ubuntu 22.0.4
O Método Profissional: Pastas Compartilhadas (Shared Folders)
Este é o melhor para quem precisa mover arquivos constantemente. Ele cria uma pasta no Windows que aparece como um "drive" ou diretório dentro do Ubuntu.
No VirtualBox, vá em Configurações da VM > Pastas Compartilhadas.
Clique no ícone de "mais" (+) e selecione a pasta no seu Windows.
Marque Apenas leitura (opcional), Montar Automaticamente e Tornar Permanente.
No Ubuntu, seu usuário precisa de permissão para ler essa pasta. Rode este comando no terminal:
Bash
sudo adduser $USER vboxsf
sudo docker start recamera
sudo docker exec -it recamera /bin/bash
========================
testes para ver como funciona chamadas no docker
docker exec recamera ls -la /workspace
sudo docker exec -it recamera bash -c "pwd && ls -la"
sudo docker exec -it recamera bash -c \
"source /workspace/tpu-mlir/envsetup.sh"
sudo docker exec -it sophgo bash -c "
source /workspace/tpu-mlir/envsetup.sh && \
pwd && \
ls -la
"
sudo docker exec -it yolo26_convert bash -c "
source /workspace/tpu-mlir/envsetup.sh
pwd
ls -la
cp hello.txt /workspace
"
sudo docker run --privileged --rm --name "sophgo" \
-v "$(pwd)/$WORK_DIR:/work" \
-w "/work/$model_name" \
sophgo/tpuc_dev:v3.1 bash -c "
source /workspace/tpu-mlir/envsetup.sh
echo '=== Step 1: Model Transform ==='
mkdir -p workspace && cd workspace
cp ../$onnx_basename .
$use_qtable && cp ../$(basename $qtable_path) . || true"
docker run --privileged -it --rm --name yolo26_convert sophgo/tpuc_dev:v3.1 bash
docker run --privileged -it --rm --name yolo26_convert \
-v "$(pwd)/$WORK_DIR:/work" \
-w "/work/$model_name" \
sophgo/tpuc_dev:v3.1 -c
"pwd"
ls
pwd
ls
echo "$(pwd)/$WORK_DIR:/work"
echo "/work/$model_name"
ls -R
"
sudo docker start yolo26_convert
sudo docker exec -it yolo26_convert /bin/bash
docker run --privileged -it --rm --name yolo26_convert sophgo/tpuc_dev:v3.1 -c "ls -R"
docker run --rm sophgo/tpuc_dev:v3.1 pwd
sudo docker run --rm \
-v "$(realpath ./$WORK_DIR):/work" \
-w "/work/yolo11n" \
sophgo/tpuc_dev:v3.1 ls -R /work
==============novo script===============
install tpu_mlir
sudo docker run -it sophgo/tpuc_dev:v3.1 bash
altere...
exit
sudo docker ps -a
sudo docker commit 16423cf57338 sophgo/tpuc_dev:v3.1-recamera
sudo docker run -it sophgo/tpuc_dev:v3.1-recamera bash
work_dir
Pasta oficial que é duplicada no lado do docker
NAO OKfaltou --chip no deploy
path test input image errado...esta como /work/tmp
deploy_args esta como "local" - remover local
case "$task" in
detect)
if [[ "$model_type" == "yolo11" ]]; then <----sempre yolo11
echo "/model.23/one2one_cv2.0/one2one_cv2.0.2/Conv_output_0,/model.23/one2one_cv3.0/one2one_cv3.0.2/Conv_output_0,/model.23/one2one_cv2.1/one2one_cv2.1>
else
adicao do -xc
665ad7751bcf
sudo docker kill 665ad7751bcf
LOG DO NOVO SCRIPT
model_transform.py --model_name yolo11n-detect --model_def best.onnx --input_shapes '[[1,3,640,640]]' --mean 0.0,0.0,0.0 --scale 0.0039216,0.0039216,0.0039216 --keep_aspect_ratio --pixel_format rgb --output_names /model.23/one2one_cv2.0/one2one_cv2.0.2/Conv_output_0,/model.23/one2one_cv3.0/one2one_cv3.0.2/Conv_output_0,/model.23/one2one_cv2.1/one2one_cv2.1.2/Conv_output_0,/model.23/one2one_cv3.1/one2one_cv3.1.2/Conv_output_0,/model.23/one2one_cv2.2/one2one_cv2.2.2/Conv_output_0,/model.23/one2one_cv3.2/one2one_cv3.2.2/Conv_output_0 --test_input /work/test.jpg --test_result yolo11n-detect_top_outputs.npz --mlir yolo11n-detect.mlir
model_runner.py --input yolo11n-detect_in_f32.npz --model best.onnx --output yolo11n-detect_ref_outputs.npz
run_calibration.py yolo11n-detect.mlir --dataset /work/dataset --input_num 100 -o yolo11n-detect_calib_table
model_deploy.py --mlir yolo11n-detect.mlir --quantize INT8 --quant_input --processor cv181x --calibration_table yolo11n-detect_calib_table --test_input /work/test.jpg --test_reference yolo11n-detect_top_outputs.npz --customization_format RGB_PACKED --fuse_preprocess --aligned_input --model yolo11n-detect_cv181x_int8.cvimodel
1 branco
2 preto
3 azul
perfect
https://www.youtube.com/watch?v=r0RspiLG260
https://www.ejtech.io/learn/train-yolo-models
https://wiki.seeedstudio.com/recamera_model_conversion/
//cars yolov8
export da sophgo
model_transform \
--model_name yolo8n \
--model_def best.onnx \
--input_shapes "[[1,3,640,640]]" \
--mean "0.0,0.0,0.0" \
--scale "0.0039216,0.0039216,0.0039216" \
--keep_aspect_ratio \
--pixel_format rgb \
--test_input cars.jpg \
--test_result yolo8n_top_outputs.npz \
--mlir yolo8n.mlir
run_calibration yolov8n.mlir \
--dataset carros \
--input_num 100 \
-o yolov8n_cali_table
model_deploy.py \
--mlir yolov8n.mlir \
--quant_input --quant_output \
--quantize INT8 \
--calibration_table yolov8n_cali_table \
--processor cv181x \
--model yolov8n_cv181x_int8_sym.cvimodel
python sample_img_object_detection.py YOLOV8 /root/cv181x/yolov8n_cv181x_int8_sym.cvimodel /root/21.jpg
obs: recamera yolov11 recamera funcionou com carros, que ultrapassam resolucao de 640x640
obs: tdl sdk parece exigir 640x640 e roda yolov8 apenas
# Manually download the specific version first
!wget https://github.com/ultralytics/assets/releases/download/v8.2.0/yolov8n.pt
# Then run your train command
!yolo task=detect mode=train model=yolov8n.pt data=data.yaml epochs=100 imgsz=640
2. Use the Python API (Recommended)
The Python API is generally less "opinionated" than the CLI and handles versioning more strictly.
Python
from ultralytics import YOLO
# This explicitly loads the V8 architecture
model = YOLO('yolov8n.pt')
# Train the model
model.train(data='data.yaml', epochs=100, imgsz=640)
--carros yolov8 -
run_calibration.py yolov8n.mlir \
--dataset ./carros
--input_num 10 \
-o yolov8n_cali_table
testes
yolov8 cars tdl_sdk
sample_img_object_detection YOLOV8 /root /root/carro.jpg
[root@milkv-duo]~# sample_img_object_detection YOLOV8 /root /root/carro.jpg
[tdl_model_factory.cpp:80] [I] input model config file is empty, load model config from /mnt/system/configs/model/model_factory.json
[tdl_model_factory.cpp:104] [I] load model config from /mnt/system/configs/model/model_factory.json done,model size:66
[tdl_model_factory.cpp:226] [I] setModelDir success,model_dir:/root
[tdl_model_factory_internal.cpp:40] [I] getModelInstance model_type:YOLOV8
[tdl_model_factory_internal.cpp:329] [I] createObjectDetectionModel success,model type:27,category:0
[tdl_model_factory.cpp:412] [I] model_path: /root/cv181x/best_cv181x.cvimodel
out_datas.size: 1
obj_meta_index: 0 class: 2 score: 0.587585 bbox: 213.596 88.3061 414.607 263.387
obj_meta_index: 1 class: 4 score: 0.893209 bbox: 488.964 35.1621 640 174.354
obj_meta_index: 2 class: 4 score: 0.844244 bbox: 447.938 189.45 639.936 366.509
obj_meta_index: 3 class: 9 score: 0.804705 bbox: 240.546 8.94987 408.378 145.323
object_meta->bboxes.size(): 4
save image to object_detection_0.jpg
[ 6911.369174] vb_cancel_block:964(): Not find pool for size(1228800).
python tdl_sophgo.py /root/cv181x/best_int8_cars.cvimodel carro.jpg
import sys
import os
from tdl import nn, image
import cv2
import numpy as np
if __name__ == "__main__":
if len(sys.argv) != 3:
print("Usage: python sample_fd.py <model_path> <image_path>")
sys.exit(1)
model_path = sys.argv[1]
img_path = sys.argv[2]
face_detector = nn.get_model(nn.ModelType.YOLOV8, model_path)
img = image.read(img_path)
# img = cv2.imread(img_path)
bboxes = face_detector.inference(img)
print(bboxes)
#https://github.com/sophgo/tdl_sdk/tree/master
python sample_img_object_detection.py YOLOV8 /root/cv181x/best_int8_cars.cvimodel carro.jpg
import sys
import os
from tdl import nn, image
import cv2
import numpy as np
def visualize_objects(img_path, bboxes, save_path="object_detection.jpg"):
"""?????????"""
img = cv2.imread(img_path)
print(f"??? {len(bboxes)} ???")
for i, obj in enumerate(bboxes):
x1, y1, x2, y2 = map(int, [obj['x1'], obj['y1'], obj['x2'], obj['y2']])
class_id = obj['class_id']
score = obj['score']
class_name = obj.get('class_name', f'class_{class_id}')
cv2.rectangle(img, (x1, y1), (x2, y2), (0, 0, 255), 2)
label = f"{class_id}:{score:.2f}"
center_x = (x1 + x2) // 2
center_y = (y1 + y2) // 2
cv2.putText(img, label, (center_x, center_y),
cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 255), 1)
cv2.imwrite(save_path, img)
print(f"?????: {save_path}")
if __name__ == "__main__":
if len(sys.argv) < 4 or len(sys.argv) > 5:
print("Usage: python3 sample_img_object_detection.py <model_id_name> <model_dir> <image_path> [threshold]")
# ???????????????
sys.exit(1)
model_id_name = sys.argv[1]
model_dir = sys.argv[2]
image_path = sys.argv[3]
threshold = float(sys.argv[4]) if len(sys.argv) == 5 else 0.5
if not os.path.exists(image_path):
print(f"???????: {image_path}")
sys.exit(1)
model_type = getattr(nn.ModelType, model_id_name)
model = nn.get_model(model_type, model_dir, device_id=0)
# ????
img = image.read(image_path)
# ????
outdatas = model.inference(img)
expected_keys = {"class_id", "class_name", "score", "x1", "y1", "x2", "y2"}
is_detection = (
isinstance(outdatas, list) and
isinstance(outdatas[0], dict) and
set(outdatas[0].keys()) == expected_keys
)
if not is_detection:
print("????????????,????:")
print(outdatas)
sys.exit(1)
print(f"out_datas.size: {len(outdatas)}")
for i, obj in enumerate(outdatas):
print(f"obj_meta_index: {i} "
f"class: {obj['class_id']} "
f"score: {obj['score']:.2f} "
f"bbox: {obj['x1']:.2f} {obj['y1']:.2f} {obj['x2']:.2f} {obj['y2']:.2f}")
visualize_objects(image_path, outdatas)
# input: python3 sample_img_object_detection.py <model_id_name> <model_dir> <image_path> [threshold]
# output: obj_meta_index: <index> class: <class_id> score: <score_value> bbox: <x1> <y1> <x2> <y2>
plates 11 recamera 100% ok
===============
model_transform.py \
--model_name yolo11n-detect \
--model_def best.onnx \
--input_shapes "[[1,3,640,640]]" \
--mean "0.0,0.0,0.0" \
--scale "0.0039216,0.0039216,0.0039216" \
--keep_aspect_ratio \
--pixel_format rgb \
--output_names "/model.23/cv2.0/cv2.0.2/Conv_output_0,/model.23/cv3.0/cv3.0.2/Conv_output_0,/model.23/cv2.1/cv2.1.2/Conv_output_0,/model.23/cv3.1/cv3.1.2/Conv_output_0,/model.23/cv2.2/cv2.2.2/Conv_output_0,/model.23/cv3.2/cv3.2.2/Conv_output_0" \
--test_input test.jpg \
--test_result yolo11n-detect_top_outputs.npz \
--mlir yolo11n-detect.mlir
run_calibration.py yolo11n-detect.mlir \
--dataset plates \
--input_num 100 \
-o yolo11n-detect_calib_table
model_deploy.py \
--mlir yolo11n-detect.mlir \
--quantize INT8 \
--quant_input \
--processor cv181x \
--calibration_table yolo11n-detect_calib_table \
--test_input test.jpg \
--test_reference yolo11n-detect_top_outputs.npz \
--customization_format RGB_PACKED \
--fuse_preprocess --aligned_input \
--model yolo11n-detect_cv181x_int8.cvimodel
plates 8 TDL_SDK 100% ok
===========================
nao esquecer fo export do site para yolo8
model_transform \
--model_name yolov8n \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0.0,0.0,0.0 \
--scale 0.0039216,0.0039216,0.0039216 \
--keep_aspect_ratio \
--pixel_format rgb \
--mlir yolov8n.mlir
run_calibration.py yolov8n.mlir \
--dataset plates \
--input_num 100 \
-o yolov8n_cali_table
python sample_img_object_detection.py YOLOV8 /root/cv181x/yolov8n_cv181x_int8_sym.cvimodel plate.jpg
/etc/init.d/S93sscma-supervisor stop
/etc/init.d/S91sscma-node stop
python sample_sergio.py /root /root/s4.jpg
===vibracao
model_transform.py \
--model_name yolo11n-detect \
--model_def best.onnx \
--input_shapes "[[1,3,640,640]]" \
--mean "0.0,0.0,0.0" \
--scale "0.0039216,0.0039216,0.0039216" \
--keep_aspect_ratio \
--pixel_format rgb \
--output_names "/model.23/cv2.0/cv2.0.2/Conv_output_0,/model.23/cv3.0/cv3.0.2/Conv_output_0,/model.23/cv2.1/cv2.1.2/Conv_output_0,/model.23/cv3.1/cv3.1.2/Conv_output_0,/model.23/cv2.2/cv2.2.2/Conv_output_0,/model.23/cv3.2/cv3.2.2/Conv_output_0" \
--test_input test_vib.jpg \
--test_result yolo11n-detect_top_outputs.npz \
--mlir yolo11n-detect.mlir
run_calibration.py yolo11n-detect.mlir \
--dataset vibracao \
--input_num 100 \
-o yolo11n-detect_calib_table
model_deploy.py \
--mlir yolo11n-detect.mlir \
--quantize INT8 \
--quant_input \
--processor cv181x \
--calibration_table yolo11n-detect_calib_table \
--test_input test_vib.jpg \
--test_reference yolo11n-detect_top_outputs.npz \
--customization_format RGB_PACKED \
--fuse_preprocess --aligned_input \
--model yolo11n-detect_cv181x_int8.cvimodel
https://universe.roboflow.com/rean-ba5ce/vibration-analysis-normal/dataset/2
yolo export model=best.pt format=onnx opset=14
AT+DEVEUI=70B3D57ED006E1A2
AT+APPEUI=0000000000000000
AT+BAND=6
AT+APPKEY=9E8B21A4624337145717626A83DA6F48
AT+MASK=0002
AT+CLASS=A
AT+JOIN=1:0:10:8
https://colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/train-yolov8-obb.ipynb#scrollTo=tdSMcABDNKW-
@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@
classificacao
C:.
└───dataset
├───train
│ ├───classe_A
│ ├───classe_B
│ └───classe_C
└───val
├───classe_A
├───classe_B
└───classe_C
!pip install ultralytics==8.2.0
!pip install ultralytics==8.3.0
from ultralytics import YOLO
import os
# 1. Desativa o W&B
os.environ["WANDB_MODE"] = "disabled"
# 2. FORÇANDO O DOWNLOAD DO MODELO DE CLASSIFICAÇÃO VIA URL
# Isso evita que ele baixe o yolo11n.pt de detecção por engano
try:
model = YOLO("yolo11n-cls.pt")
except:
# Se ele insistir no 11, usamos o link direto do 8-cls
model = YOLO("https://github.com/ultralytics/assets/releases/download/v8.2.0/yolo11n-cls.pt")
# 3. Treino
# Aponte para a pasta que contém 'train' e 'val'
model.train(
data="/content/dataset",
epochs=100,
imgsz=640,
task='classify'
)
from google.colab import drive
drive.mount('/content/drive')
e então
cp /content/drive/MyDrive/buggio/dataset.zip /content/.
ou
!yolo task=classify mode=train \
model=yolov8n-cls.yaml \
data=/content/dataset \
epochs=100 \
imgsz=640
yolo export model=best.pt format=onnx imgsz=640
model_transform.py \
--model_name yolov8n \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0.0,0.0,0.0 \
--scale 0.0039216,0.0039216,0.0039216 \
--keep_aspect_ratio \
--pixel_format rgb \
--mlir yolov8n.mlir
run_calibration.py yolov8n.mlir \
--dataset motors \
--input_num 100 \
-o yolov8n_cali_table
model_deploy.py \
--mlir yolov8n.mlir \
--quant_input --quant_output \
--quantize INT8 \
--calibration_table yolov8n_cali_table \
--processor cv181x \
--model yolov8n_cv181x_int8_sym.cvimodel
[16:23, 12/02/2026] Sergio milk-v Duo S: edita o model_factory.json que fica em /mnt/system/configs/model
[16:24, 12/02/2026] Sergio milk-v Duo S: "YOLOV8":{
"_comment":"custom model, specify num_cls or branch string",
"file_name":"yolov8n_cv181x_int8_sym_4"
},
[16:25, 12/02/2026] Sergio milk-v Duo S: só edita o "file_name": com o nome do seu modelo (lembranso que sem o _cv181x no final, pois ele coloca automatico)
python sample_img_classification.py /root/cv181x/yolov8n_cv181x_int8_sym.cvimodel /root/bom_c0.jpg
import os
import cv2
import time
import subprocess
import numpy as np
from pinpong.libs.dfrobot_lis2dh import LIS2DH
from pinpong.board import Board, I2C
# Funcao para configurar pinos no Milk-V Duo
def execute_shell_command(command):
try:
result = subprocess.run(command, shell=True, capture_output=True, text=True)
return result.stdout if result.stdout else result.stderr
except Exception as e:
return str(e)
# Inicializacao do Hardware
Board("milkv-duo").begin()
execute_shell_command('duo-pinmux -w B20/IIC4_SCL')
execute_shell_command('duo-pinmux -w B21/IIC4_SDA')
lis = LIS2DH()
while not lis.init(lis.LIS2DH12_RANGE_8GA):
print("Aguardando sensor LIS2DH...")
time.sleep(1)
# Configuracao do Dataset
DATASET_DIR = "dataset_yolo"
CLASSES = [] # Lista de classes capturadas
def preparar_pastas():
# Cria estrutura: dataset_yolo/train/images e dataset_yolo/train/labels
for folder in ["images", "labels"]:
path = os.path.join(DATASET_DIR, "train", folder)
if not os.path.exists(path):
os.makedirs(path)
def criar_yaml():
# Gera o arquivo de configuracao para o YOLO11
yaml_path = os.path.join(DATASET_DIR, "data.yaml")
with open(yaml_path, "w") as f:
f.write(f"train: {os.path.abspath(DATASET_DIR)}/train/images\n")
f.write(f"val: {os.path.abspath(DATASET_DIR)}/train/images\n")
f.write("\nnc:\n")
f.write(f"names: {CLASSES}\n")
print(f"\n[OK] Arquivo data.yaml criado em: {yaml_path}")
def capturar_ciclo(label_nome, class_id, quantidade=100):
print(f"\nIniciando captura de {quantidade} amostras para: {label_nome}")
img_dir = os.path.join(DATASET_DIR, "train", "images")
lbl_dir = os.path.join(DATASET_DIR, "train", "labels")
for i in range(quantidade):
amostras = []
for _ in range(256):
lis.read_XYZ()
amostras.append([lis.x, lis.y, lis.z])
time.sleep(0.001) # Amostragem rapida
# Processamento dos dados (Escala 8G = +/- 8000mg)
dados = np.array(amostras)
dados = np.clip(dados, -8000, 8000)
imagem_base = ((dados + 8000) * 255 / 16000).astype(np.uint8)
# Criacao da imagem 256x256 RGB (OpenCV usa BGR)
# img_final[linha, coluna, canal]
img_final = np.zeros((256, 256, 3), dtype=np.uint8)
for linha in range(256):
# Atribuindo 256 cores baseadas nos eixos para os 3 canais
img_final[linha, :, 0] = imagem_base[linha, 2] # Canal Blue (Eixo Z)
img_final[linha, :, 1] = imagem_base[linha, 1] # Canal Green (Eixo Y)
img_final[linha, :, 2] = imagem_base[linha, 0] # Canal Red (Eixo X)
# Nomes dos arquivos
file_base = f"{label_nome}_c{i}"
# Salva Imagem em JPG com qualidade maxima (100) para preservar dados
img_path = os.path.join(img_dir, f"{file_base}.jpg")
cv2.imwrite(img_path, img_final, [int(cv2.IMWRITE_JPEG_QUALITY), 100])
# Salva Label (YOLO format: class_id x_center y_center width height)
with open(os.path.join(lbl_dir, f"{file_base}.txt"), "w") as f:
f.write(f"{class_id} 0.5 0.5 1.0 1.0")
if i % 20 == 0:
print(f"Gravando: {i}/{quantidade}...")
# --- EXECUCAO DOS 3 CICLOS ---
preparar_pastas()
for ciclo in range(3):
print(f"\n--- CICLO {ciclo + 1} DE 3 ---")
nome_classe = input("Digite o nome para esta vibracao (ex: normal, erro, alta): ").strip()
if nome_classe not in CLASSES:
CLASSES.append(nome_classe)
id_atual = CLASSES.index(nome_classe)
input(f"Prepare o sensor para '{nome_classe}' e pressione ENTER para capturar 100 imagens...")
capturar_ciclo(nome_classe, id_atual, 100)
criar_yaml()
print("\nDataset concluido com sucesso!")
yolo export model=best.pt format=onnx opset=14
======vibracao===============
=bom sempre usar test_input
yolo8n
=============================
yolo export model=best.pt format=onnx opset=14
model_transform \
--model_name yolo8n \
--model_def best.onnx \
--input_shapes "[[1,3,640,640]]" \
--mean "0.0,0.0,0.0" \
--scale "0.0039216,0.0039216,0.0039216" \
--keep_aspect_ratio \
--pixel_format rgb \
--test_input 2bal640x640.jpg \
--test_result yolo8n_top_outputs.npz \
--mlir yolo8n.mlir
run_calibration yolov8n.mlir \
--dataset vibracao \
--input_num 100 \
-o yolov8n_cali_table
model_deploy.py \
--mlir yolov8n.mlir \
--quant_input --quant_output \
--quantize INT8 \
--calibration_table yolov8n_cali_table \
--processor cv181x \
--test_input 2bal640x640.jpg \
--model yolov8n_cv181x_int8_sym.cvimodel
====vibracao
==solucao tdl sdk...(foi usado)
python sample_img_object_detection.py YOLOV8 /root/cv181x/yolov8n_cv181x_int8_sym_cv181x.cvimodel /root/4ballfaultserious_ke-48_png.rf.e5badf73f45eeb9d25224c8f7cb40d13.jpg
run_calibration.py yolov8n.mlir \
--dataset hz \
--input_num 100 \
-o yolov8n_cali_table
model_deploy.py \
--mlir yolov8n.mlir \
--quant_input --quant_output \
--quantize INT8 \
--calibration_table yolov8n_cali_table \
--processor cv181x \
--model yolov8n_cv181x_int8_sym.cvimodel
picolm
tinyllama-1.1b-chat-v1.0-q4_k_m.gguf
./picolm /mnt/usb/llm/tinyllama-1.1b-chat-v1.0-q4_k_m.gguf -p "The meaning of life is" -n 100
De-me 5 ideias de nomes para uma loja de cafe?
====
para o blog...e tdl sdk...
new colab teste - chocolates...
https://colab.research.google.com/github/EdjeElectronics/Train-and-Deploy-YOLO-Models/blob/main/Train_YOLO_Models.ipynb#scrollTo=EMEDk5byzxY5
!pip install ultralytics==8.3.0 #o segredo para o yolo nao pegar .pt superior
já tenha uma pasta /test
from ultralytics import YOLO
model = YOLO('yolov8n.pt')
results = model.train(data=f"data.yaml", epochs=100, imgsz=640)
!yolo detect predict model=runs/detect/train2/weights/best.pt source=/content/train/test save=True
import glob
from IPython.display import Image, display
for image_path in glob.glob(f'/content/runs/detect/predict/*.jpg')[:10]:
display(Image(filename=image_path, height=400))
print('\n')
https://www.youtube.com/watch?si=frhm3ByszGu_lBmT&v=r0RspiLG260&feature=youtu.be
https://colab.research.google.com/
model_deploy.py \
--mlir yolo11n-detect.mlir \
--quantize BF16 \
--quant_input \
--processor cv181x \
--calibration_table yolo11n-detect_calib_table \
--test_input test.jpg \
--test_reference yolo11n-detect_top_outputs.npz \
--customization_format RGB_PACKED \
--fuse_preprocess --aligned_input \
--model yolo11n-detect_cv181x_int8.cvimodel
https://ricardoadulis.sharepoint.com/:u:/s/smartcore/IQC5Hj8OhqzLSahmy_Qu2jFYASIYqhr7HpmA0Wsre_sWRU0?e=RN3dU9
https://github.com/sophgo/tdl_sdk/blob/master/tool/yolo_export/yolov8_export.py
https://milkv.io/docs/duo/getting-started/setup
https://platform.ultralytics.com/home
//ultra web platform
yolo26n
model_transform \
--model_name yolo26n \
--model_def best.onnx \
--input_shapes "[[1,3,640,640]]" \
--mean "0.0,0.0,0.0" \
--scale "0.0039216,0.0039216,0.0039216" \
--keep_aspect_ratio \
--pixel_format rgb \
--test_input test.jpg \
--test_result yolo26n_top_outputs.npz \
--mlir yolo26n.mlir
run_calibration \
yolo26n.mlir \
--dataset cores \
--input_num 100 \
-o yolo26n_calib_table
model_deploy \
--mlir yolo26n.mlir \
--quantize INT8 \
--quant_input \
--processor cv181x \
--calibration_table yolo26n_calib_table \
--test_input test.jpg \
--test_reference yolo26n_top_outputs.npz \
--customization_format RGB_PACKED \
--fuse_preprocess \
--aligned_input \
--model yolo26n_1684x_int8_sym.cvimodel
model_deploy \
--mlir yolo26n.mlir \
--quantize BF16 \
--quant_input \
--processor cv181x \
--calibration_table yolo26n_calib_table \
--test_input test.jpg \
--test_reference yolo26n_top_outputs.npz \
--customization_format RGB_PACKED \
--fuse_preprocess \
--aligned_input \
--model yolo26n_1684x_int8_sym.cvimodel
https://universe.roboflow.com/tyler-yonjx/car-colors-1smyc
\\wsl.localhost\docker-desktop\mnt\docker-desktop-disk\data\docker\volumes\6e0475e8fbeba5b744b0d593cbfd6ad76a5f70300caec98595e193b6110b9dd9
++++++++++++++++++
classificacao tdl-sdk (yolov8)
++++++++++++++++++
https://colab.research.google.com/github/roboflow-ai/notebooks/blob/main/notebooks/train-yolo11-object-detection-on-custom-dataset.ipynb#scrollTo=tdSMcABDNKW-yolo11s
!pip install ultralytics
criar pasta dataset e copiar zip dentro
treine
data.yaml
path: /content/train
train: train
val: valid
test: test
!yolo task=classify mode=train \
model=yolov8n-cls.yaml \
data=/content/dataset \
epochs=100 \
imgsz=640
ou
!yolo task=classify mode=train \
model=yolov8n-cls \
data=/content/dataset \
epochs=100 \
imgsz=640
ou
!yolo classify train data=/content/dataset model=yolov8n-cls.pt epochs=50 imgsz=640
yolo export model=best.pt format=onnx imgsz=640
model_transform \
--model_name yolov8n-cls \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0.0,0.0,0.0 \
--scale 0.0039216,0.0039216,0.0039216 \
--keep_aspect_ratio \
--pixel_format rgb \
--mlir yolov8n.mlir
run_calibration.py yolov8n.mlir \
--dataset cores \
--input_num 100 \
-o yolov8n_cali_table
model_deploy.py \
--mlir yolov8n.mlir \
--quant_input --quant_output \
--quantize INT8 \
--calibration_table yolov8n_cali_table \
--processor cv181x \
--model yolov8n_cv181x_int8_sym.cvimodel
run_calibration.py yolov8n.mlir \
--dataset cores \
--input_num 100 \
-o yolov8n_cali_table
sample_img_classification CLS_YOLOV8 /root /root/test.jpg
python sample_img_classification.py yolov8n_cv181x_int8_sym.cvimodel test.jpg
!yolo classify train data=/content/dataset model=yolov8n-cls.pt epochs=50 imgsz=640
TENTAR...
model_transform \
--model_name yolov8_cls \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0,0,0 \
--scale 0.0039216,0.0039216,0.0039216 \
--mlir yolov8_cls.mlir
run_calibration.py yolov8_cls.mlir \
--dataset cores \
--input_num 100 \
-o yolov8n_cali_table
model_deploy.py \
--mlir yolov8_cls.mlir \
--quant_input --quant_output \
--quantize INT8 \
--calibration_table yolov8n_cali_table \
--processor cv181x \
--model yolov8n_cv181x_int8_sym.cvimodel
========blog
treinei no ultralytics platform
usei o python export_class
model_transform.py \
--model_name yolo_cls \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0.0,0.0,0.0 \
--scale 0.0039216,0.0039216,0.0039216 \
--pixel_format bgr \
--keep_aspect_ratio \
--channel_format nchw \
--mlir yolo_cls.mlir
run_calibration.py yolo_cls.mlir \
--dataset cores \
--input_num 100 \
-o yolo_cls_cali_table
model_deploy.py \
--mlir yolo_cls.mlir \
--quantize INT8 \
--calibration_table yolo_cls_cali_table \
--fuse_preprocess \
--asymmetric \
--processor cv181x \
--model yolo_cls_cv181x_int8_sym.cvimodel
python sample_img_classification.py /root/yolo_cls_cv181x_int8_sym.cvimodel /root/test.jpg
sample_img_classification CLS_YOLOV8 /root /root/test.jpg
====desisto?
ultralytics platform
convert_class.py
model_transform.py \
--model_name yolov8n_cls \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean 0.0,0.0,0.0 \
--scale 0.0039216,0.0039216,0.0039216 \
--keep_aspect_ratio \
--pixel_format rgb \
--test_input test.jpg \
--test_result yolov8n_cls_top_outputs.npz \
--mlir yolov8n_cls.mlir
run_calibration.py yolov8n_cls.mlir \
--dataset cores \
--input_num 100 \
-o yolo_cls_cali_table
model_deploy.py \
--mlir yolov8n_cls.mlir \
--quantize INT8 \
--calibration_table yolo_cls_cali_table \
--processor cv181x \
--test_input test.jpg \
--test_reference yolov8n_cls_top_outputs.npz \
--tolerance 0.85,0.45 \
--fuse_preprocess \
--customization_format RGB_PLANAR \
--aligned_input \
--model yolo_cls_cv181x_int8_sym.cvimodel
???yolo task=classify export model=best.pt format=onnx???
./sample_img_classification /root CLS_YOLOV8 /root/test.jpg
sample_img_classification /root/yolo_cls_cv181x_int8_sym.cvimodel CLS_YOLOV8 /root/test.jpg
https://pt.aliexpress.com/item/1005009136395783.html?ug_edm_item_id=1005009136395783&edm_click_module=alg_product_3_4586567300&creative_img_ind=7&edm_log_data=gmod-edm-item-list-three-columns.track-edm-item-list-three-columns-log-link&tracelog=rowan&rowan_id1=aeug_edm_41180_1_pt_BR_2026-03-03&rowan_msg_id=tcpiCOWS_41180_%24b76fd4f35ee8490f8be62bee8ab63495&ck=in_edm_other&mem_info=ODmQN4iVZeMljqCr%2Fd0gpw%3D%3D-100003-aeug_edm_41180-IaxZ1cvIAFXa2co01j1vpwMJ553K%20YAzAjqcUK4jAsE%3D&gatewayAdapt=glo2bra#nav-specification
Nícolas Zimermann - Detecção e Controle de Acesso veicular com Milk-v
Selmo Werner Júnior - Sistema de Monitoramento Térmico para Manutenção Preditiva Industrial Utilizando Computação de Borda em Arquitetura com Milk-V
Artur Lian Fernandes Torres - Desenvolvimento de um Dispositivo para Monitoramento de Condição de Máquinas Rotativas Utilizando Análise de Frequência e Rede Neural Convolucional com Milk-V
./sample_img_classification CLS_HAND_GESTURE /root /root/test.jpg
./sample_img_classification CLS_HAND_GESTURE /root /root/ruido_200Hz_200Hz_c1.jpg
python sample_img_classification.py /root/cv181x/cls_hand_gesture_128_128_INT8_cv181x.cvimodel /root/test.jpg
python sample_img_classification.py /root/cv181x/cls_hand_gesture_128_128_INT8_cv181x.cvimodel /root/ruido_200Hz_200Hz_c1.jpg
=================================================
TESTE 1.27 TPU MLIR
python sample_img_classification_.py /root/cv181x/yolo_cls_cv181x_int8_sym.cvimodel /root/ruido_200Hz_200Hz_c1.jpg
./sample_img_classification CLS_YOLOV8 /root /root/ruido_200Hz_200Hz_c1.jpg
melhorou...mas...
model_transform.py \
--model_name yolov8n_cls \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mlir yolov8n_cls.mlir
run_calibration.py yolov8n_cls.mlir \
--dataset cores \
--input_num 100 \
-o yolo_cls_cali_table
model_deploy.py \
--mlir yolov8n_cls.mlir \
--quantize INT8 \
--quant_input \
--calibration_table yolo_cls_cali_table \
--processor cv181x \
--model yolo_cls_cv181x_int8_sym.cvimodel
==================================================
./sample_img_classification CLS_YOLOV8 /root /root/test.jpg
=*****************************=
model_transform.py \
--model_name yolov8n_cls \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean "0,0,0" \
--scale "0.00392156862745098,0.00392156862745098,0.00392156862745098" \
--keep_aspect_ratio \
--pixel_format rgb \
--channel_format nchw \
--test_input test.jpg \
--tolerance 0.99,0.99 \
--test_result yolov8n_cls_top_outputs.npz \
--mlir yolov8n_cls.mlir
run_calibration.py yolov8n_cls.mlir \
--dataset cores \
--input_num 100 \
-o yolo_cls_cali_table
model_deploy.py \
--mlir yolov8n_cls.mlir \
--quantize INT8 \
--quant_input \
--calibration_table yolo_cls_cali_table \
--processor cv181x \
--test_input yolov8n_cls_in_f32.npz \
--test_reference yolov8n_cls_top_outputs.npz \
--tolerance 0.9,0.6 \
--model yolo_cls_cv181x_int8_sym.cvimodel
\\\\\\\\\\\\\\\\\\\\\\\
model_transform.py \
--model_name yolov8n_cls \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean "0,0,0" \
--scale "0.00392156862745098,0.00392156862745098,0.00392156862745098" \
--keep_aspect_ratio \
--pixel_format rgb \
--channel_format nchw \
--output_names /model.9/linear/Gemm_output_0 \
--test_input test.jpg \
--test_result yolov8n_cls_top_outputs.npz \
--mlir yolov8n_cls.mlir
run_calibration.py yolov8n_cls.mlir \
--dataset cores \
--input_num 100 \
-o yolo_cls_cali_table
model_deploy.py \
--mlir yolov8n_cls.mlir \
--quantize INT8 \
--quant_input \
--calibration_table yolo_cls_cali_table \
--processor cv181x \
--test_input yolov8n_cls_in_f32.npz \
--test_reference yolov8n_cls_top_outputs.npz \
--tolerance 0.9,0.6 \
--model yolo_cls_cv181x_int8_sym.cvimodel
=====hub maix
model_transform.py \
--model_name yolov8n_cls \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean "0,0,0" \
--scale "0.017124753831663668, 0.017124753831663668, 0.017124753831663668" \
--keep_aspect_ratio \
--pixel_format rgb \
--channel_format nchw \
--output_names /model.9/linear/Gemm_output_0 \
--test_input test.jpg \
--test_result yolov8n_cls_top_outputs.npz \
--mlir yolov8n_cls.mlir
model_transform.py \
--model_name yolov8n_cls \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean "0,0,0" \
--scale "0.00392156862745098, 0.00392156862745098, 0.00392156862745098" \
--keep_aspect_ratio \
--output_names "output0" \
--pixel_format rgb \
--channel_format nchw \
--test_input green.jpg \
--test_result yolov8n_cls_top_outputs.npz \
--tolerance 0.95,0.60 \
--mlir yolov8n_cls.mlir
run_calibration.py yolov8n_cls.mlir \
--dataset testes \
--input_num 1600 \
-o yolo_cls_cali_table
model_deploy.py \
--mlir yolov8n_cls.mlir \
--quantize INT8 \
--quant_input \
--calibration_table yolo_cls_cali_table \
--processor cv181x \
--test_input yolov8n_cls_in_f32.npz \
--test_reference yolov8n_cls_top_outputs.npz \
--tolerance 0.95,0.60 \
--model yolo_cls_cv181x_int8_sym.cvimodel
Look this https://wiki.sipeed.com/maixpy/doc/en/ai_model_converter/maixcam.html
NAO ESQUECER DO BATCH QUE REMOVE SOFTMAX E SAIDAS = OUTPUT0
"mean":[123.5, 123.5, 123.5],
"std":[255.0,255.0,255.0]
python sample_img_classification_.py /root/yolo_cls_cv181x_int8_sym.cvimodel /root/ruido_200Hz_200Hz_c1.jpg
----monkey
model_transform.py \
--model_name yolov8n_cls \
--model_def best.onnx \
--input_shapes [[1,3,640,640]] \
--mean "0,0,0" \
--scale "0.00392156862745098, 0.00392156862745098, 0.00392156862745098" \
--keep_aspect_ratio \
--output_names "output0" \
--pixel_format rgb \
--channel_format nchw \
--test_input Ades_1-2.jpg \
--test_result yolov8n_cls_top_outputs.npz \
--tolerance 0.95,0.60 \
--mlir yolov8n_cls.mlir
run_calibration.py yolov8n_cls.mlir \
--dataset train \
--input_num 500 \
-o yolo_cls_cali_table
model_deploy.py \
--mlir yolov8n_cls.mlir \
--quantize INT8 \
--quant_input \
--calibration_table yolo_cls_cali_table \
--processor cv181x \
--test_input yolov8n_cls_in_f32.npz \
--test_reference yolov8n_cls_top_outputs.npz \
--tolerance 0.95,0.60 \
--model yolo_cls_cv181x_int8_sym.cvimodel
./sample_img_classification CLS_YOLOV8 /root /root/barbie2.jpg
python sample_img_classification_.py /root/yolo_cls_cv181x_int8_sym.cvimodel /root/barbie2.jpg
[
{
"id": "5665c4bbc82b4729",
"type": "function",
"z": "8c5f3c081178a986",
"name": "function 1",
"func": "if (!msg.payload.data.keypoints.length) return null;\n\nlet kp = msg.payload.data.keypoints[0][1];\nlet conf = 50;\n\nlet estado = null;\n\n// detectar braço levantado\nif (kp[9][2] > conf && kp[9][1] < kp[5][1]) estado = \"RIGHT HAND\";\nif (kp[10][2] > conf && kp[10][1] < kp[6][1]) estado = \"LEFT HAND\";\n\n// estado anterior\nlet ultimo = context.get(\"ultimo_estado\");\n\n// só envia se mudou\nif (estado !== ultimo) {\n\n context.set(\"ultimo_estado\", estado);\n\n if (estado !== null) {\n msg.payload = estado;\n return msg;\n }\n}\n\nreturn null;",
"outputs": 1,
"timeout": 0,
"noerr": 0,
"initialize": "",
"finalize": "",
"libs": [],
"x": 160,
"y": 480,
"wires": [
[
"615a8a0791045d32"
]
]
}
]
if (!msg.payload.data.keypoints.length) return null;
let kp = msg.payload.data.keypoints[0][1];
let conf = 50;
let estado = null;
// detectar braço levantado
if (kp[9][2] > conf && kp[9][1] < kp[5][1]) estado = "LEFT HAND";
if (kp[10][2] > conf && kp[10][1] < kp[6][1]) estado = "RIGHT HAND";
// estado anterior
let ultimo = context.get("ultimo_estado");
// só envia se mudou
if (estado !== ultimo) {
context.set("ultimo_estado", estado);
if (estado !== null) {
msg.payload = estado;
return msg;
}
}
return null;
Nenhum comentário:
Postar um comentário